Fuzzy classifier design (Q5955187)

The aim of this book is to provide a structured view into the pattern classification methods that use fuzzy sets at some stage of the design of the classifier. Chapter 1 exposes some of the pros and the cons in the debate on whether or not fuzzy sets provide a useful tool for designing pattern classifiers. The data sets used for illustration throughout the book are introduced. Chapter 2 is a brief, reference-like detour through the classics of statistical pattern recognition and Bayes decision theory. Chapter 3 details several approaches to statistical classifier design. Parametric and nonparametric classifiers are derived from the Bayesian classifier model. Finding prototypes for the \(k\)-nearest neighbor and nearest prototype classifier is a special accent of this chapter. Three popular neural network models are introduced: multi-layer perceptron, radial-basis function network and learning vector quantization networks. Chapter 4 introduces fuzzy set theory to the extent that is needed for understanding fuzzy classifier designs thereafter. The emphasis is on basic operations on fuzzy sets, especially fuzzy aggregation methods. Practical issues such as determining the membership functions are also discussed. Chapter 5 explains how fuzzy if-then systems work. The Mamdani-Assilian and Takagi-Sugeno-Kang (TSK) models are explained and translated into pattern classifiers. The last section of Chapter 5 investigates some theoretical properties of fuzzy if-then models. The universal approximation property of fuzzy TSK systems is revisited with respect to the pattern classification task. A caveat is indicated: fuzzy if-then classifiers could be simple look-up tables in disguise. Various options for training of fuzzy if-then classifiers are explored in Chapter 6. While some of them are only sketched, (e.g., using neuro-fuzzy models), others are explained in more details (e.g., min-max or hyperbox designs) with numerical examples and experiments. Chapter 7 presents non if-then fuzzy classifiers. Many such models appeared in the 80s but were then overrun by the more successful if-then stream. Some early models are summarized at the beginning of the chapter. The two most successful non-if-then designs are outlined next: fuzzy \(k\)-nearest neighbor and fuzzy prototype classifiers. Ten fuzzy \(k\)-nn variants are tested with the four data sets used in the book (two synthetic sets, Satimage and Phoneme). The generalized nearest prototype classifier is introduced as a common framework for a number of fuzzy and non-fuzzy classifier models. Combination of multiple classifiers is discussed in Chapter 8. Various fuzzy and non-fuzzy schemes for classifier fusion and classifier selection are described, 28 of which are also illustrated experimentally. Majority vote over dependent classifiers is analyzed on a synthetic example. The target audience of this book are academic researchers, graduate and postgraduate students in mathematics, engineering, computer science and related disciplines.